AI agents are useful

I think it’s very cool that you can use Codex as a personal assistant that actually does things for you.

I usually use the CLI on Linux, but I’ve been testing the desktop application on Windows. One thing I like is being able to ask it to open a website and update something for me. If I’m not already logged in, I can log in and then let it continue. It feels like having a personal assistant who can take care of those small tasks that otherwise keep getting postponed.

For instance, maybe you have a profile page somewhere and want to update the description about yourself. You can ask it to help write the text and then update the page for you. With the browser capabilities available in the desktop app, that kind of workflow becomes quite convenient.

I’ve also tried it on a Mac with DaVinci Resolve Studio. It now includes support for connecting AI assistants, so you can ask an assistant to do things inside the application. I think these kinds of integrations make agents much more useful. They can work in the applications you already use.

What interests me especially is all the things I wouldn’t otherwise bother doing. You have profile pages on so many websites. Do you really care enough to keep all of them up to date? Or maybe there are accessibility improvements you would like to make to a website, but you never find the time.

I have two children, and I need to prioritize. Like most people, I’m quite busy during the day. If something takes an hour or more, it might never happen. But if it takes a few minutes of prompting and then checking that the result looks right, suddenly it becomes worth doing. That changes the calculation.

Recently, I read about United Airlines restricting unauthorized AI agents from booking and managing flights. In a recent announcement, United pointed to the risks of sharing account credentials, unintended purchases or account changes, and regulatory requirements. One example it gave was the human acknowledgment during check-in that your bags don’t contain prohibited items.

Still, I can understand why people want an AI to book their tickets. Anyone who has booked a flight knows how frustrating it can be. There are different travel agencies with different prices, sometimes different from booking directly with the airline. Then there are all the pages about baggage, seats, contact details, passport details, and payment. It’s a lot of boring and tedious steps. I might try getting an agent to help the next time I need to book a ticket.

OpenClaw attracted a lot of attention earlier this year, and it feels like there has been growing interest in using agents for everyday tasks since then.

Of course, there have also been security problems. Even with these problems, the technology is already very useful. Once people get used to having help with these tasks, I find it hard to imagine them wanting to go back to doing everything manually.

And the models are only part of it. The harnesses are improving too. So are the applications for managing agents.

For instance, Herdr provides a workspace for running and managing agents, including a terminal interface that can serve as an alternative to a tmux-based workflow. Tools like that make it easier to work with several agents in parallel.

There are also more competitors than OpenAI and Anthropic. Moonshot AI has Kimi Code, and DeepSeek has been developing its own agent harness. There are open-source alternatives such as OpenCode as well. Different companies and projects are building both the models and the software around them. That’s one reason I think this will continue even if some of the companies struggle.

I’m a bit skeptical about the valuations in the AI industry and how the spending will work out. I also wonder how much today’s pricing reflects what these services will eventually need to charge. If using agents becomes much more expensive, that could change the calculation again.

But I can be skeptical about the business side and still find the technology useful. If OpenAI or Anthropic runs into problems, there are other companies and open-source projects that can pick up the gauntlet. I don’t think the whole thing depends on those two companies succeeding.

It reminds me a bit of the dot-com bubble. People could see that the internet would become useful for shopping and many other things, but some investors overestimated how quickly that would turn into profitable businesses.

There is a similar comparison with railway expansion. A lot of railway companies went bankrupt, but railways were still useful, and we still have them today. A useful technology doesn’t automatically make every company building it a good investment.

Investors want a return, and sometimes their expectations about how long that will take don’t match reality. I think that could happen with AI too. There could be a bubble popping, companies disappearing, or a period when people become much less enthusiastic.

But I expect people to keep using the tools that save them time. From what I’ve tried already, there are enough useful things here that I would miss them if they disappeared.